Comparing Computational Cognitive Models of Generalization in a Language Acquisition Task

Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 96–106, Austin, Texas, November 1-5, 2016. c⃝2016 Association for Computational Linguistics Comparing Computational Cognitive Models of Generalization in a Language Acquisition Task Libby Barak, Adele E. Goldberg, Psychology Department Princeton University Princeton, NJ, USA {lbarak,adele}@princeton.edu Suzanne Stevenson Department of Computer Science University of Toronto Toronto, Canada suzanne@cs.toronto.edu Abstract Natural language acquisition relies on appro- priate generalization: the ability to produce novel sentences, while learning to restrict pro- ductions to acceptable forms in the language. Psycholinguists have proposed various prop- erties that might play a role in guiding appro- priate generalizations, looking at learning of verb alternations as a testbed. Several com- putational cognitive models have explored as- pects of this phenomenon, but their results are hard to compare given the high variability in the linguistic properties represented in their input. In this paper, we directly compare two recent approaches, a Bayesian model and a connectionist model, in their ability to repli- cate human judgments of appropriate gener- alizations. We find that the Bayesian model more accurately mimics the judgments due to its richer learning mechanism that can exploit distributional properties of the input in a man- ner consistent with human behaviour. 1 Introduction Native speakers of a language are mostly able to generalize appropriately beyond the observed data while avoiding overgeneralizations. A testbed area for studying generalization behavior in language ac- quisition is verb alternations – i.e., learning the pat- terns of acceptability of alternative constructions for expressing similar meanings. For example, English speakers readily use a new verb like text in both the double-object (DO) construction (“text me the de- tails”) and the prepositional-dative (PD) (“text the details to me”) – an instance of the dative alterna- tion. However, speakers avoid overgeneralizing the DO construction to verbs such as explain that resist its use (“?explain me the details”), even though they occur with analogous arguments in the PD alterna- tive (“explain the details to me”). Psycholinguis- tic studies have focused on the possible properties of natural language that enable such generalization while constraining it to acceptable forms. Initially, children are linguistically conservative: they generally use verbs in constructions that are very close to exemplars in the input (Lieven et al., 1997; Akhtar, 1999; Tomasello, 2003; Boyd and Goldberg, 2009). Children reach adult-like com- petence by gradually forming more general as- sociations of constructions to meaning that allow them to extend verb usages to unwitnessed forms. Much work has emphasized the role of verb classes that capture the regularities across semantically- similar verbs, enabling appropriate generalization (e.g., Pinker, 1989; Fisher, 1999; Levin, 1993; Am- bridge et al., 2008). Usage-based approaches have argued that such class-based behaviour can arise in learning through the clustering of observed usages that share semantic and syntactic properties (e.g., Bybee, 2010; Tomasello, 2003; Goldberg, 2006). A number of studies also reveal that the statistical properties of the language play a central role in lim- iting generalization (e.g., Bresnan and Ford, 2010; Ambridge et …

Content truncated to 500 words